[patch 05/39] genirq/msi: Remove filter from msi_free_descs_free_range()

[Thomas Gleixner]

Miquel!

On Wed, Mar 01 2023 at 11:55, Miquel Raynal wrote:
> tglx at linutronix.de wrote on Fri, 11 Nov 2022 14:54:22 +0100 (CET):
>
>> When a range of descriptors is freed then all of them are not associated to
>> a linux interrupt. Remove the filter and add a warning to the free function.
>> +		/* Leak the descriptor when it is still referenced */
>> +		if (WARN_ON_ONCE(msi_desc_match(desc, MSI_DESC_ASSOCIATED)))
>> +			continue;
>> +		msi_free_desc(desc);
>>  	}
>>  }
>
> It looks like since this commit I am getting warnings upon EPROBE_DEFER
> errors in the mvpp2 Marvell Ethernet driver. I looked a bit at the
> internals to understand why this warning was shown, and it seems that
> nothing "de-references" the descriptors, which would mean here:
> resetting desc->irq to 0.

Correct. This platform-msi ^(*&!@&^ hack really needs to die ASAP.

Marc, where are we on that? Is this still in limbo?

> I am wondering how useful thisd WARN_ON() is, or otherwise where the

It is useful as it caught bugs already.

> desc->irq entry should be zeroed (if I understand that correctly), any
> help will be appreciated.

Untested workaround below. I hate it with a passion, but *shrug*.

Thanks,

        tglx
---
 drivers/base/platform-msi.c |    1 +
 include/linux/msi.h         |    2 ++
 kernel/irq/msi.c            |   23 ++++++++++++++++++++++-
 3 files changed, 25 insertions(+), 1 deletion(-)

--- a/drivers/base/platform-msi.c
+++ b/drivers/base/platform-msi.c
@@ -324,6 +324,7 @@ void platform_msi_device_domain_free(str
 	struct platform_msi_priv_data *data = domain->host_data;
 
 	msi_lock_descs(data->dev);
+	msi_domain_depopulate_descs(data->dev, virq, nr_irqs);
 	irq_domain_free_irqs_common(domain, virq, nr_irqs);
 	msi_free_msi_descs_range(data->dev, virq, virq + nr_irqs - 1);
 	msi_unlock_descs(data->dev);
--- a/include/linux/msi.h
+++ b/include/linux/msi.h
@@ -631,6 +631,8 @@ int msi_domain_prepare_irqs(struct irq_d
 			    int nvec, msi_alloc_info_t *args);
 int msi_domain_populate_irqs(struct irq_domain *domain, struct device *dev,
 			     int virq, int nvec, msi_alloc_info_t *args);
+void msi_domain_depopulate_descs(struct device *dev, int virq, int nvec);
+
 struct irq_domain *
 __platform_msi_create_device_domain(struct device *dev,
 				    unsigned int nvec,
--- a/kernel/irq/msi.c
+++ b/kernel/irq/msi.c
@@ -1109,14 +1109,35 @@ int msi_domain_populate_irqs(struct irq_
 	return 0;
 
 fail:
-	for (--virq; virq >= virq_base; virq--)
+	for (--virq; virq >= virq_base; virq--) {
+		msi_domain_depopulate_descs(dev, virq, 1);
 		irq_domain_free_irqs_common(domain, virq, 1);
+	}
 	msi_domain_free_descs(dev, &ctrl);
 unlock:
 	msi_unlock_descs(dev);
 	return ret;
 }
 
+void msi_domain_depopulate_descs(struct device *dev, int virq_base, int nvec)
+{
+	struct msi_ctrl ctrl = {
+		.domid	= MSI_DEFAULT_DOMAIN,
+		.first  = virq_base,
+		.last	= virq_base + nvec - 1,
+	};
+	struct msi_desc *desc;
+	struct xarray *xa;
+	unsigned long idx;
+
+	if (!msi_ctrl_valid(dev, &ctrl))
+		return;
+
+	xa = &dev->msi.data->__domains[ctrl.domid].store;
+	xa_for_each_range(xa, idx, desc, ctrl.first, ctrl.last)
+		desc->irq = 0;
+}
+
 /*
  * Carefully check whether the device can use reservation mode. If
  * reservation mode is enabled then the early activation will assign a